|
[08 Jun 2015|01:48am] |
|
this entry is backdated
|
|
| n-Rooks / simplex algorithm / bipartite matching |
[09 May 2008|12:16am] |
Someone has proposed a better solution to my "constrained n-Rooks" problem that I don't understand.
"constrained n-Rooks": return every possible assignment of rooks to squares on an n x n chessboard, under the constraint that no two rooks threaten each other, and that there are some given squares in which rooks can't be placed.
<< Define a graph with edge weights of zero or one, depending on whether or
not there is an edge. Bipartite matching can be written as an LP.
You find some solution to the problem. Now you want to find the other
optimal solutions. These solutions form a connected set in the polytope,
so you can just to breadth first search with coloring from any starting solution,
only stepping through other optimal solutions. >>
If you understand how bipartite matching relates to my problem, how it can be solved as above, and can explain this to me by this weekend, I'll buy you dinner.
I do not understand the Simplex algorithm.
|
|
| buying flights early |
[07 May 2008|07:46pm] |
Knowing that a lot of people will soon be buying tickets from America and Canada to Helsinki on July 3, should I expect to get a better deal if I buy early? Are the airlines aware of conferences like ICML (~500 people)?
Tangentially, Farecast doesn't work outside of America.
|
|
| paper style |
[07 May 2008|07:06pm] |
The first section of my paper is not titled "Introduction", because I want the title to be informative. I cannot think of any other papers violating this regularity.
Likewise, the titles of my sections and subsections tend to be long. Seems like a good alternative to having a "table of contents, in paragraph format and without page numbers (yechh)".
--
I'm also very fond of footnotes, it seems. There's normally a trade-off between accuracy/completeness vs. conciseness/clarity. Footnotes let me have both (by apologizing for the inaccuracy/incompleteness in the main text).
|
|
| Voice Post |
[07 May 2008|03:40pm] |
|
|
|
| sick, part two |
[07 May 2008|10:40am] |
Last night, my fever reached 103. Strangely, my only symptoms have been the fever, chills, thermal discomfort, etc. and fatigue (standing up was difficult), and a little bit of joint swelling (which I infer from a very slight pain): no coughing, no mucus, no stomach issues. And I only sneezed once.
A good friend came by to look after me. Since this has been going on for several hours, we considered going to the emergency room. I was not able to locate information of which hospitals my insurance will pay for. Apparently, if you're admitted into the ER, it's always covered. But this was a fuzzy line.
I also called the 24h doctor on duty, and she took 1 hour and 35 minutes to return my call. The operators kept saying it would be 10 minutes. Also, urgent care clinics in Pittsburgh inaccessible to those who don't have a car. (good luck getting a taxi!)
Call me spoiled, but I'm not used to this level of service. In Brazil, you can just show up at the doctor's office, and if it's late at night, we rely on family and friends who are doctors.
I've also been having strange thoughts during my "sleep": the same words kept cycling over and over again. Napkin... serviette... napkin... serviette.
Today, I have a doctors appointment at 3:30 p.m. I'm not better yet.
|
|
| sick |
[06 May 2008|04:41pm] |
I am suddenly very very tired, with a high fever and my only option for being seen by a doctor today is to go to the ER. This sucks.
Maybe this was going around at Cornell.
I hit my head pretty good on Thursday evening (bottom of the swimming pool), but passed the concussion test (which I asked to get). I hope that's not related.
|
|
| accepted papers |
[05 May 2008|07:39pm] |
UAI: 32 plenary talks, 40 posters (total 72)
ICML, total: 156
COLT, total: 44
MLG seems interesting too.
4 conferences in one!
|
|
| algorithmic problem |
[05 May 2008|04:14pm] |
Given a set L of permutations of a list, compute the maximal partial order (DAG) under which all lists in L are topological sorts. How would you do this? Can you do better than O(m n^2 k^2)? (m = number of lists in L, k = # of things in domain = the size of the elements of L)
Sounds like an ILP problem.
One idea: whenever a cycle is detected, all the elements are deemed incomparable.
|
|
| NESCAI |
[01 May 2008|11:33am] |
I'm going to Cornell tomorrow, mostly for the networking.
I should make a poster tonight.
|
|
| MediaWiki LaTeX still not working |
[30 Apr 2008|09:33pm] |
Also, the File Upload seems to want to write to the wrong place:
The upload directory (public) is not writable by the webserver.
I suspect they are related.
I can't believe how much time I've spent on this MediaWiki LaTeX issue. Maybe 15 hours in total. I just want to friggin pay someone to fix this for me, but I don't know where to find them.
|
|
| proof style |
[29 Apr 2008|03:21pm] |
|
I'd really like to write "end suppose" in the camera-ready version of my paper.
|
|
| Helsinki |
[28 Apr 2008|06:11pm] |
My collaborators got me a studio on a different campus of the university. I'm also getting $1000, which I hope can pay for the flights (although currently, my best deal is $1228). Then there's registration fees: 650Euro, unless I can claim to be a student by the time I register (which I can't), in which case 365Euro.
---
<< The apartments are fully furnished with linen, cutlery and television. A HOASnet Intenet connection and cleaning twice a month are included in the rent. The cleaning doesn't include washing dishes and taking out the garbage. There is a laundry room free of charge in the building. Tenants may reserve a sauna hour in Hoas Viikki local office situated in Cubile (open Mon-Fri 1 pm to 4 pm). The sauna is located in the student dormitory just across the street. There is a parking place free of charge - please notify Hoas of your car registration number to avoid parking tickets. In all matters except the sauna reservations and parking, please contact Hoas Kamppi office. >>
---
I also found out that "plenary talk" means the whole conference will be watching me (my future employers, colleagues, etc.)! I'll want to practice it many times. Peter thinks it should be 25 minutes long with 5 minutes for questions. This is some serious incentive to try to run my method on real data... if I can find it! (Unfortunately, it seems economists usually publish summaries of their data, which make parametric assumptions)
In 2006, the proceedings published 72 papers. They had 26 plenary talks and 42 posters. (4 absentees?)
In 2007, there were 26 plenary talks (+ 4 invited talks) and 31 posters (that's 57 papers).
Also: << Note that all papers appear as full length papers in the UAI proceedings. The only difference between poster and plenary papers is the type of presentation at the conference.>>
In any case, it's exciting and scary!
|
|
| gossip from mathland |
[28 Apr 2008|12:29pm] |
Proof theory is good at translating infinitary results (ergodic theory) into finitary ones (combinatorics), thanks to Kohlenbach's revolution of the field (few structural proof theorists remain). Terry Tao is interested and published a paper related to this.
source: ![[info]](http://p-stat.livejournal.com/img/userinfo.gif) htowsner
|
|
| Busy Beaver, and adding axioms about non-halting |
[26 Apr 2008|07:49pm] |
The value of Σ is not known for n >= 5, because there are machines that we haven't been able to run long enough to halt or prove that they never halt (also, there is no hope of proving that they halt if we use a small axiomatic system: Google "ten pounds of axioms" for philosophical controversies).
Lacking evidence to the contrary, could it be justified to assume that a particular TM never halts? (note that many of these *deserve* to be axioms, since they are true and don't follow from our small axiom sets; and if we are wrong, we *will* eventually find out)
Such assumptions would allow us to solve open Pi_1 problems (like the Goldbach Conjecture) that can be expressed in n bits where Σ(n) is known, i.e. the assumption is that Σ(n) isn't bigger than the current champion. (the argument is that KC(1st counterexample) <= KC(statement). Thus 1st counterexample <= Σ(n) where n is the length of the statement)
This seems like a fine working assumption, analogous to axioms used in the natural sciences.
I can imagine proceeding this way, by choosing axioms that best explain mathematical regularities (including unproven ones), perhaps in a Bayesian way. Having said that, even if our assumption is falsified (if the TM is shown to halt afterall), might it be useful to continue pretending otherwise? (do axioms that take a long time to be refuted tend to be, for application purposes, less significant?)
Chaitin said that we need to be empirical about axioms, and that's what I'm suggesting here.
|
|
| Gaussian Processes |
[26 Apr 2008|05:57pm] |
Gaussian Processes have been on my mind since Wednesday. Here are some thoughts:
GPs must obey certain constraints:
* correlation function must yield a valid correlation matrix (be symmetric, obey the triangle inequality for correlation)
* no "quantum entanglement", i.e. observing leads to a normal Bayesian update: the function P(X1) = SUM_x2 P(X2=x2)P(X1|X2=x2)
One can encode a linear trend by adding a linear term to the function.
What about:
* circular domains
* periodic domains: can we encode more correlations? What does the correlation between peaks say about the Fourier Transform?
|
|
| my plans for July |
[25 Apr 2008|10:18am] |
I'm going to be in Helsinki for July, between roughly 4 and 28. I'm afraid that those will turn out to be the longest days of my life.
Not only was my paper accepted as a plenary talk, but all kinds of people have been offering to help me pay for it.
First of all, my collaborators from Helsinki offered to pay for my accommodation, given that I'll be there for two weeks after the conference.
Secondly, my local supervisor said that they can come up with money to pay a large chunk of my expenses (perhaps more).
Thirdly, my boss from Machine Learning, who is totally unrelated to this project, offered to chip in a couple hundred bucks.
Now, given that I'll probably fly out on July 3, I probably do not want to have a connection inside the United States.
I'm going to have 2.5 free weekends in Helsinki. Here are some travel ideas:
* visit St. Petersburg (I may have visa issues)
* visit the far north of the country, to see the midnight sun.
* visit Stockholm, in case Henrik is there
My other life requirements should be easy to arrange, though I'll probably pay for them myself:
* Internet
* cell phone (one would expect that, from Finland)
* bicycle
|
|
| Google and identity |
[23 Apr 2008|10:40pm] |
Is there some way for me to label Google's top hits for "Gustavo Lacerda" as "me" or "not me" in a way that will help future searches?
www.optimizelife.com (me)
www.gustavolacerda.com.br (not me)
gustavolacerda.livejournal.com (me)
www.cliki.net/Gustavo%20Lacerda (me)
www.cs.cmu.edu/~wcohen/pubs.html (me)
www.informatik.uni-trier.de/~ley/db/indices/a-tree/l/Lacerda:Gustavo.html (me, for now)
www.amazon.com/gp/pdp/profile/A8G1C4VYGK5GJ (me)
www.nabble.com/MozDev---annozilla-f11274.html (me)
fare.livejournal.com/data/foaf (me)
douweosinga.com/projects/languageevolution (me)
staff.science.uva.nl/~kamps/publications/coauthors.html (me)
breyten.livejournal.com/data/foaf (me)
lists.tunes.org/archives/tunes/2002-November/003456.html (me)
penelopebeolchi.blogspot.com/2008/03/ensaio-thas-moll-e-gustavo-lacerda.html (not me)
adsoftheworld.com/taxonomy/industry/other?page=2 (not me)
It really can't be very hard to learn a good classifier here, by using bag-of-words and link structure.
|
|
